AI founders have long expected fast-moving competitors. The more immediate strategic threat may now be the foundation-model provider underneath their product.
TechCrunch is framing that issue in an upcoming Disrupt 2026 Builders Stage session, “What Happens When OpenAI Ships Your Roadmap,” scheduled for October 13–15 in San Francisco. The event session will feature Airbyte co-founder and CEO Michel Tricot, Webflow CEO Linda Tong, and Radical Ventures partner Rob Toews.
The premise is straightforward: as OpenAI, Anthropic and Google regularly add capabilities, products built around a single model capability can be recast as a feature in a platform update. For operators, the question is no longer only whether a team can build a useful AI product. It is whether that product retains a reason to exist when the underlying platform improves.
The change: model capability is becoming less ownable
Foundation-model advances can rapidly lower the cost of functions that once required specialized AI engineering. That raises the bar for startups whose proposition is primarily a thin layer over a model, such as a standalone interface, prompt workflow or generalized automation.
This does not mean every application company is doomed to commoditization. It means the source of differentiation matters more. A company that sells access to a broadly available capability has limited protection if the platform provider bundles the same capability. A company that makes that capability useful inside a difficult business process has a more durable starting point.
That distinction also changes how founders should describe the business to customers and investors. “We use the best model” is not a moat when customers can increasingly access comparable models and tools elsewhere.
Where defensibility can still sit
The source material identifies several areas that model providers cannot easily replicate: proprietary data, embedded workflows, customer relationships, domain expertise and trust. Those are not interchangeable advantages, and each creates different operating work.
Data can be defensible when a company has legitimate, ongoing access to data that improves outcomes in a specific use case—not merely a one-time corpus or generic public information.
Workflow integration matters when software is connected to systems of record, approvals, permissions, exceptions and accountability. Replacing it then requires more than switching to a newer model endpoint.
Domain expertise and trust are especially relevant in high-consequence settings, where users need outputs that fit established processes and can be reviewed, governed and acted upon. The model may generate an answer; the company still needs to make that answer operationally reliable.
Airbyte’s position illustrates the infrastructure side of this equation. TechCrunch says its open-source data-integration platform serves more than 7,000 customers, including 18% of the Fortune 500. That type of installed workflow and data connectivity is a different asset from a single model feature.
What operators should do now
Treat major foundation-model launches as recurring strategic tests. For each announced capability, ask three questions:
1. What part of our product has become cheaper or easier to reproduce? Be explicit about the feature, not just the broad category. 2. What customer outcome remains difficult without our software, integrations and expertise? This is the value that should shape product investment. 3. Could a customer move to the platform directly without disrupting a critical workflow? If the answer is yes, the company has a retention and positioning problem to solve.
Product roadmaps should reserve capacity for model substitution, evaluation and redesign. Reliance on one provider can be commercially useful, but it also means a provider’s release calendar can dictate application-company priorities. Teams should separate the model-dependent layer from the parts that encode customer workflow, data controls and product logic wherever practical.
What to watch next
The meaningful signal is not simply whether model companies release more features; that is expected. Watch where those features move from general capability into packaged workflows, enterprise controls and distribution channels that overlap with application vendors.
Also watch how buyers evaluate AI vendors. As baseline capabilities become more accessible, purchasing decisions should increasingly turn on implementation speed, integration depth, governance, reliability and measurable business outcomes.
For founders, the practical aim is not to predict every platform release. It is to ensure that every release strengthens the product’s economics or customer outcome rather than erasing its reason for being.
